Realtime RGB-Based 3D Object Pose Detection Using Convolutional Neural Networks

We present a new approach to efficiently detect the 3D pose of objects in images. The approach uses a single neural network we call TQ-Net to predict the translation vector and the quaternion, after which the quaternion is converted to a rotation matrix, and through a projection algorithm, we can project eight points surrounding the object to the image and connect them to form a 3D bounding box. Considering there is a constraint existing in the quaternion, we add a normalization layer to get a more precise result. Experiments show that our approach performs well and can detect 3D poses in real time. Our approach processes images with a resolution of 1280*720 at 61 frames per second on a GTX 1080 GPU.